Predicting how a rail buckles

Paper: Wongkaew, Muanyoksakul, Ngamkhanong, Sresakoolchai, Kaewunruen. Data driven machine learning prognostics of buckling failure modes in ballasted railway track. Discover Applied Sciences 6, 212 (April 2024). PDF on this site: buckling_01.pdf. Project notes: here.

This was the first journal paper. The question was not “will the rail fail in heat?” Railway engineers already know continuous welded rail can buckle. The question was which failure, on which geometry, early enough to change maintenance.

Question

Two modes matter in the field. Snap-through is sudden lateral jump. Progressive is a slower walk-out, often at lower lateral stiffness and lower temperature. They are not the same repair, and they are not the same speed restriction. A binary “buckle / no buckle” classifier would have been easier and less useful.

Climate makes the question operational in Thailand: hotter days, CWR, ballast that is not a European textbook.

Data and method

Labels came from finite-element simulations, not from waiting for a pile-up. About 8,000 track scenarios. The features the model was allowed to see:

• lateral stiffness
• displacement limits
• torsional resistance
• unconstrained length
• initial misalignment

Outputs: non-buckling, snap-through, progressive.

Most simulated tracks do not fail, so a dummy that always says “no buckle” looks good. We used SMOTE on the minority modes and k-fold validation. Models in the bake-off: logistic regression, kNN, decision trees, random forest, XGBoost, LightGBM. XGBoost was the one we kept. On that evaluation, F1 was 0.97. Feature importance put lateral misalignment, torsional resistance, and lateral displacement limit at the top.

We then took the trained model to a real alignment in Thailand and asked whether it recovered mode and the temperature neighbourhood of the failure. The paper and the poster are the place for the case plots. I will not restyle them here.

What I would not do again

Misalignment in the FEM set was a sine wave. Real track is a mess of welds and tamps. The paper already flags that. Next time I want measured geometry in the training distribution, not only in the case study. And I would report the confusion between snap-through and progressive as the headline metric, not overall F1 — overall F1 is too easy to inflate with the majority class even after SMOTE.

The habit this paper taught me, which the method post is about: pick a label a maintainer already has a word for, then make the ML answer in that word.